Currently, Reconstruct 2D Classes only outputs 2D classes, and does not output the matching particle sets.
In certain situations, it would be useful to be able to use the classes out of Reconstruct 2D for selection (e.g. ran 2D with very low maximum resolution for speed; reconstruct 2D to generate interpretable classes. Or ran 2D on subtracted particles, but want to select in the context of the unsubtracted particle).
Most recently I encountered this when testing 2D classification on denoised particles, where I was classifying on denoised, then reconstructing from re-extracted non-denoised particles using alignments & class assignments from denoised - the 2D classes & class fidelity are surprisingly good. The idea of this is to use 2D alignments from denoised particles to improve particle centering, then perform further processing with non-denoised particles.
Currently, if I queue a select 2D job with the input particles and the classes from reconstruct 2D, it gives nonsense results - the classes don’t match the correct number of particles.
Cheers
Oli
(PS - it would also sometimes be useful to allow multiple inputs for class 2D, so that when 2D is run on subsets of data, classes can be selected in a single Select2D job)
Hey @olibclarke – very cool to hear that classifying images from denoised micrographs is showing promise!
I just ran a basic pipeline of class2D → reconstruct 2D → select 2D (with classes from reconstruct and particles from class2D) and all the classes have the appropriate size and are matched up appropriately. Could you give us a bit more info about the workflow here and any non-default params?
Thanks!
Valentin
Hi Valentin,
Class2D (denoised particles) :
Re-extract from non-denoised mics, then reconstruct2D & select2D:
You can see the class assignments are all messed up.. and I input 250k particles but only 35k are showing in select2D
Or another case - Select 2D, reconstruct 2D, select 2D:
1 Like
Ah ok, that second example made everything quite clear! You’re totally right that this currently broken – noted and tracked! (we’ll likely do exactly as you suggest: amend reconstruct 2D to also output particles).
1 Like
One other thing re reconstruct 2D - it currently lacks the “hard classify” option that is present in class 2D. This might be handy to add for parity - often I don’t run Class2D to the last iteration, as I will set it with more final iterations than I need so I can monitor convergence and kill/complete it once done. It would be useful to be able to test if hard-classify is useful, by running reconstruct 2D on a prior 2D classification run of the same dataset.
It would also be useful to have a “re-align classes” option - so that similar views are shown in similar orientations, even if the “similarity” sorting method was not enabled in Class2D
2 Likes
